CN-122019189-A - Cluster CPU cooperative frequency adjustment method and system, computer equipment and medium

CN122019189ACN 122019189 ACN122019189 ACN 122019189ACN-122019189-A

Abstract

The present disclosure provides a cluster CPU cooperative frequency adjustment method and system, a computer device and a medium. The cluster CPU cooperative frequency adjustment method comprises the steps of generating global total frequency adjustment parameters through a trained time sequence regression model, synchronizing the global total frequency adjustment parameters to all nodes of a whole cluster by adopting a seed relay diffusion mode based on an adjacent topological relation among the nodes, updating an adjacent frequency adjustment table locally maintained by each node, completing distributed optimal frequency adjustment representative node election based on the adjacent frequency adjustment table after all the nodes of the whole cluster are synchronized, executing inter-node distribution of the global total frequency adjustment parameters through the selected representative nodes, generating target operation frequencies of each node, and executing CPU frequency adjustment operation based on the distributed target operation frequencies. Embodiments of the present disclosure are capable of achieving a coordinated frequency of multiple nodes in the near future of loading tides.

Inventors

YU SHENGJIN
YUE LONGGUANG
MO QINGLIANG

Assignees

联通数字科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. The cluster CPU cooperative frequency adjustment method is characterized by comprising the following steps of: Generating a predicted global total frequency modulation parameter through a trained time sequence regression model; Based on the adjacent topological relation among the nodes, synchronizing the global total frequency modulation parameter to all the nodes of the whole cluster by adopting a seed relay diffusion mode, and updating an adjacent frequency modulation table locally maintained by each node; Based on the adjacent frequency modulation table after all nodes of the whole cluster are synchronized, finishing the representative node election of the distributed optimal frequency modulation, and generating the target running frequency of each node by executing the inter-node distribution of the global total frequency modulation parameter by the selected representative node; each node performs a CPU frequency adjustment operation based on the assigned target operating frequency.
2. The method of claim 1, wherein the step of determining the position of the substrate comprises, The input of the trained time sequence regression model is the global total CPU load time sequence of the whole cluster, the unique output is the global total frequency modulation parameter, and the global total frequency modulation parameter is the sum of the total CPU equivalent frequencies required to be provided by all nodes of the whole cluster in a future scheduling period.
3. The method of claim 2, wherein prior to generating the predicted global total tuning parameters by the trained time series regression model, the method further comprises: Constructing the time sequence regression model by adopting a cyclic neural network model; Performing supervised training based on historical cluster load timing data of the global total CPU load timing sequence to obtain the trained timing regression model, The generating predicted global total tuning parameters by the trained time series regression model further comprises: and generating a system load predicted value in a future time period through the trained time sequence regression model, and outputting the global total frequency modulation parameter based on the system load predicted value.
4. The method of claim 1, wherein synchronizing the global total fm parameter to all nodes of a full cluster using a seed relay diffusion scheme further comprises: Using a node generating global total frequency modulation parameters as a seed node, and transmitting frequency modulation prior information packets to the direct connection 1-hop adjacent node of the seed node; In response to receiving the frequency modulation prior information packet, the adjacent node updates a local adjacent frequency modulation table, and relay forwards the hop count of the frequency modulation prior information packet to the adjacent node of the adjacent node after adding one; And when the hop count of the frequency modulation prior information packet reaches a preset maximum threshold value or the receiving node has processed the frequency modulation prior information packet, terminating the information forwarding flow of the corresponding branch of the node.
5. The method of claim 1, wherein performing the distributed optimal tuning representative node election based on the synchronized adjacency list of all nodes of the full cluster, and performing the inter-node allocation of the global total tuning parameter by the elected representative node further comprises: calculating a frequency modulation capability score of each node based on state data in the local adjacency frequency modulation table of each node; Based on the adjacent topological relation between the nodes, a main representative node is generated by electing in a distributed pairwise consensus mode; Dividing the cluster into at least one continuous adjacent area based on a topological adjacency principle, generating a unique area representative node of each continuous adjacent area through distributed voting in the continuous adjacent area, wherein the number of the continuous adjacent areas is a square root downward integer value of the total node number of the cluster; And the main representative node executes the distribution of the global total frequency modulation parameter, and each continuous adjacent area representative node completes the distribution and the distribution of the single node frequency in the corresponding continuous adjacent area.
6. The method of claim 5, wherein the step of determining the position of the probe is performed, The regional representative node is used for summarizing node state data in a corresponding continuous adjacent region and issuing adjacent relay of a frequency modulation strategy, wherein the regional representative node is selected from candidate regional nodes which meet the constraint conditions that the online state is normal, the CPU utilization rate does not exceed a preset threshold, the hardware frequency modulation function is normal and at least one direct adjacent node is provided; the master representative node performing the allocation of global total tuning parameters further comprises: calculating single node reference frequency, wherein the single node reference frequency is the quotient of the global total frequency modulation parameter and the total number of normal nodes of the cluster; The method comprises the steps of carrying out weighted distribution on residual quota after global total frequency modulation parameters are deducted from a guard quota, and determining total frequency quota of each continuous adjacent area, wherein the guard quota is the minimum cooperation of the continuous adjacent area meeting the minimum performance requirement of service, and the weight is the proportion of average load of each continuous adjacent area to total load of the whole cluster; Transmitting the total frequency quota of each continuous adjacent area to the representative node of the corresponding continuous adjacent area through seed diffusion; The region representing node allocates single node frequencies based on node loads and hardware states in corresponding contiguous regions.
7. The method of claim 1, wherein the performing CPU frequency adjustment operations further comprises: each node reads the allocated target operating frequency from the local adjacent frequency modulation table, and calls the kernel frequency modulation interface to carry out frequency adjustment by expanding the Berkeley data packet filter; In response to the end of frequency, each node updates the local node state data in the local adjacency list and synchronizes to the directly connected adjacency node through adjacency heartbeat broadcast, and/or Each node in the cluster independently maintains an adjacent frequency modulation table in a local kernel shared memory, wherein the adjacent frequency modulation table comprises a global anchoring area, a local node state area, a direct-connection adjacent node state area, a diffusion adjacent node state area and a global decision area, and the adjacent frequency modulation table is used for storing corresponding data of the whole frequency modulation process.
8. A clustered CPU co-frequency adjustment system, the clustered CPU co-frequency adjustment system comprising a plurality of nodes, at least one of the plurality of nodes loaded with a trained time series regression model, the clustered CPU co-frequency adjustment system configured to: Generating predicted global total frequency modulation parameters through the trained time sequence regression model; Based on the adjacent topological relation among the nodes, synchronizing the global total frequency modulation parameter to all the nodes of the whole cluster by adopting a seed relay diffusion mode, and updating an adjacent frequency modulation table locally maintained by each node; Based on the adjacent frequency modulation table after all nodes of the whole cluster are synchronized, finishing the representative node election of the distributed optimal frequency modulation, and generating the target running frequency of each node by executing the inter-node distribution of the global total frequency modulation parameter by the selected representative node; each node performs a CPU frequency adjustment operation based on the assigned target operating frequency.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that, The processor, when executing the program, implements a clustered CPU coordination frequency adjustment method as defined in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program, when executed by a processor, implements a clustered CPU coordination frequency adjustment method as defined in any one of claims 1 to 7.

Description

Cluster CPU cooperative frequency adjustment method and system, computer equipment and medium Technical Field The present disclosure relates to the field of communications. And more particularly, to a cluster CPU cooperative frequency adjustment method, a cluster CPU cooperative frequency adjustment system, a computer device, and a computer-readable storage medium. Background With the development of cloud computing and high performance computing, servers are faced with the current situation of severely fluctuating loads and unpredictable challenges to operating system scheduling, where conventional task schedulers (CFS, EEVDF) and central processing unit (Central Processing Unit, CPU) frequency modulation strategies rely primarily on detection reaction mechanisms. That is, the load state of the current system is monitored, so that the CPU frequency or other operations can be improved, however, certain hysteresis exists, and when the burst flow arrives, the CPU frequency rise and the core wake-up delay cause jitter or action delay in the initial response stage of the service. However, if the load data is predicted and tuned in advance for a single node in the cluster CPU, the calculation amount is excessive and the control redundancy is caused in practical application. Disclosure of Invention The invention aims to provide a cluster CPU cooperative frequency adjustment method, which comprises the following steps: Generating a predicted global total frequency modulation parameter through a trained time sequence regression model; based on the adjacent topological relation among the nodes, synchronizing the global total frequency modulation parameters to all the nodes of the whole cluster by adopting a seed relay diffusion mode, and updating an adjacent frequency modulation table locally maintained by each node; Based on the adjacent frequency modulation table after all nodes of the whole cluster are synchronized, finishing the representative node election of the distributed optimal frequency modulation, and generating the target running frequency of each node by executing the inter-node distribution of the global total frequency modulation parameter by the selected representative node; each node performs a CPU frequency adjustment operation based on the assigned target operating frequency. Optionally, the input of the trained time sequence regression model is the global total CPU load time sequence of the whole cluster, the unique output is the global total frequency modulation parameter, and the global total frequency modulation parameter is the sum of the total CPU equivalent frequencies required to be provided by all nodes of the whole cluster in one scheduling period in the future. Optionally, generating the predicted global total tuning parameter by the trained time series regression model further comprises: Constructing a time sequence regression model by adopting a cyclic neural network model; Performing supervised training based on historical cluster load timing data of the global total CPU load timing sequence to obtain a trained timing regression model, Generating the predicted global total tuning parameters by the trained time series regression model further comprises: And generating a system load predicted value in a future time period through the trained time sequence regression model, and outputting a global total frequency modulation parameter based on the system load predicted value. Optionally, synchronizing the global total frequency modulation parameter to all nodes of the full cluster by adopting a seed relay diffusion mode further comprises: using the node generating the global total frequency modulation parameter as a seed node, and transmitting a frequency modulation prior information packet by the seed node to only the direct connection 1-hop adjacent node; In response to receiving the frequency modulation prior information packet, the adjacent node updates a local adjacent frequency modulation table, and relay forwarding is carried out to the adjacent node after adding one hop count of the frequency modulation prior information packet; And when the hop count of the frequency modulation prior information packet reaches a preset maximum threshold value or the receiving node has processed the frequency modulation prior information packet, terminating the information forwarding flow of the corresponding branch of the node. Optionally, based on the adjacency list after all nodes of the whole cluster are synchronized, completing the election of the representative nodes of the distributed optimal frequency modulation, and performing the inter-node allocation of the global total frequency modulation parameter by the selected representative nodes further comprises: calculating a frequency modulation capability score of each node based on state data in the local adjacency frequency modulation table of each node; Based on the adjacent topological relation between the nodes, a main representative n